AI-Assisted Adaptive Monitoring for Blockchain-Based Food Quality Assurance and Supply Chain Risk Detection
Main article
Abstract
Modern food supply chains generate continuous, high-volume sensor streams that are expensive to record on a public ledger and frequently made redundant during stable operating conditions. Yet the same chains must remain auditable, with critical events such as cold chain breaks or microbial spoilage detected and immutably documented. This paper proposes an AI-assisted adaptive monitoring framework that couples a permission blockchain with edge-deployed machine learning models to record only what matters, when it matters. A long short-term memory (LSTM) network estimates a real-time risk score from environmental and process variables, while a context score derived from logistics phase, product perishability, and historical incidence weights the importance of each measurement. The combined score drives a three-mode adaptive sampling controller that scales sensing rate, edge aggregation, and on-chain anchoring to the present level of risk. A Hyperledger Fabric prototype with smart contracts for state transitions, hash anchoring of off-chain bulk streams, and stakeholder-specific access policies provides trust substrate. We evaluate the framework on a synthetic but realistic 30-day dairy cold-chain workload comprising approximately 1,100 shipment-days. Compared with a 0.10 Hz static baseline, the proposed approach reduces transmitted sensor volume by 87.4 % over a 24-hour window and reduces day-30 cumulative on-chain storage by 99.9 % versus full on-chain logging and 48.5 % versus a static-threshold event policy, while improving the F1-score of risk-event detection from 0.641 to 0.927 and reducing mean detection latency from 62.0 s to 11.2 s. The framework supports SDG 9 and SDG 12 by enabling resource-efficient, transparent food traceability across distributed multi-actor environments.
